What a Distilled Model Inherits From Its Teacher

Take a heavily censored Chinese model, distill its outputs into GPT-OSS, and does the censorship come along? The finding here is that it does not, while financial reasoning gets sharper rather than duller. Reasoning scores rise noticeably and cost per query lands well below the comparison models, which are the two results put forward. The self-distillation method and the evaluation protocol are laid out alongside, and the invitation is to reproduce it rather than take the numbers on faith.